Two years into the current AI cycle, enterprise results are increasingly bimodal. A small set of use-cases are compounding into structural margin gains; a much larger set are absorbing engineering capacity and producing dashboards no one reads.
The argument
Applied AI creates durable margin in exactly three places: where it removes a repetitive judgement that was previously bottlenecked on a scarce expert, where it compresses a cycle time that customers were willing to pay a premium to avoid, and where it makes a decision that was previously made by gut feel auditable. Everything else is interesting and most of it is not commercial.
What we see in the field
The pattern repeats across our engagements: the use-cases that earn out have a named owner inside the P&L they are supposed to improve, a baseline measurement that pre-dates the AI work, and a fallback path the business can use when the model is wrong. The use-cases that quietly fail have a sponsor in a central function, an aspiration in place of a baseline, and no human-readable behaviour when confidence drops.
What it changes
The bimodal outcome means the value of an AI programme is increasingly determined by what it refuses to fund, not what it funds. Boards comfortable killing well-presented pilots are now structurally ahead of boards that aren't.
Where to start
Run a single hour of triage on the current AI portfolio against three questions: who owns the P&L line this is supposed to move, what was the baseline before we started, and what is the safe behaviour when the model is unsure. Use-cases without an answer to all three should be paused.
